Using machine learning to detect bots in World of Warcraft Creating a decision tree based on play styles to classify between a human and a program By Ian Stevens Supervisor: Wolfgang Mayer
What is a bot A game bot is an automated program that is designed to preforms certain tasks with in a game. Game bots are mostly used for: levelling up a character, Farming certain items, or/and making a task easier for a player. Popular bot programs in Wo. W include honer buddy and Lazy bot.
How bots ruin the game Players that use bots to level up misrepresent their proficiency in the game. Take all the item drops away for paying player. Gives players an unfair advantage. Ruin the in game economy.
Earlier forms of bots People farming items and selling them for real money Prison camps in china. Made prisoners farm items. Other forms of bot ◦ Social media ◦ Chat bots
How they go undetected User can change the setting of the bot to tell it how to act. Game bot companies update their product to counter act new bot detection methods.
Current bot detection short falls All bot detection method used currently suffer for one short fall or another. ◦ Increases the latency of the game. ◦ Ruin the emersion of the game. ◦ Invade users privacy, current Wo. W bot detection system.
The aim of my thesis Develop a bot detection method that ◦ avoids the short falls in current bot detection methods. ◦ Make it hard for a company to counter this detection method by focusing on aspects on game play easily over looked. ◦ Have a low computational cost. Try to record and process as few factors as possible.
Factors of game play Percentage of Items Gathered. Distance between the player and item drop. when farming. Enemies engaged. Time between item collections. Which drops are gathered.
Main Research questions How effective is this method at detecting bots? How much does each factor contribute to classifying a Wo. W avatar? Can bots be identified with fewer factors? How well does it classify bot profile outliers? ◦ What percentage of bots are detected ? ◦ How does increasing the training size effect the outcome? ◦ How many humans are classified as bots?
Mining the data Percentage of Items Gathered/Distance between player and item drop when farming. ◦ The location of the player. ◦ The location of all item in the area. Enemies engaged ◦ The location of enemies. Which items were gathered. ◦ When the player is farming. Time between item collection. ◦ The Time. This information is saved every 0. 5 seconds to a data base.
The participants Advanced players: Players who have levelled up multiple characters to level 80. Intermediate players: Players who have played multiple hours of Wo. W but just levelled up a single character to level 80. Beginner players: Players who haven’t played much/any Wo. W but know the basics.
Conducting the experiment Each participant will be asked to perform four ten minute farming sessions starting at different location. They will play on a server I set up with Jeuties Repack. They will all use the same character. They will be told to farm a specific item, Saronite ore.
The bot programs Lazybot and Honorbuddy. 72 test cases in total, 10 min each. Three different routes. two different “Do not approach target with certain number of enemies around” number. two different combat approaches, casting different spells. Each route will be tested three times starting at a different location.
Classifier C. 4. 5 decision tree in weka 1 st test 2 nd test 3 rd test ◦ Trained with 20 bot and 9 human cases. ◦ Test with 52 bot and 39 human cases. ◦ Trained with 30 bot and 18 human cases. ◦ Test with 42 bot and 30 human cases. ◦ Repeat 1 st and 2 nd trained ◦ Test with 30 new bot profiles, a lot of which are outliers
Test with less factors Drop the factor with the least information gain. Repeat all three tests, compare results.
Looking at other methods Online game bot detection based on party -play log analysis. Identified 95. 92% of bot cases. ◦ Bots don’t always form parties. A Behaviour Analysis-Based game bot detection approach considering various play Styles. Identified 95. 35% of bot cases. ◦ High computational cost.
Future work Test on a lager scale. Test how easy it is develop a bot to go unnoticed. If bad results, add more factors.
Acknowledgments Wolfgang Mayer, Supervisor. The people who will take part in the study. Stuart Hadfield, helped with setting up the server and bot programs.
Скачать презентацию Using machine learning to detect bots in World
1bff7a9b5e9fa50193b81fde10a702b5.ppt